Identifying druggable pockets in proteins is central to structure-based drug discovery, yet conventional ligand-binding-site prediction and cryptic-pocket prediction are typically treated as separate tasks requiring distinct tools, preprocessing pipelines, and evaluation protocols. This separation is limiting because the two problems share a common biological foundation: co-evolutionary patterns encoded in protein sequences carry implicit signals about both conventional and cryptic binding sites. We present UniPocket, a unified multitask framework for residue-level prediction of both ligand-binding and cryptic-pocket residues within a single shared-backbone architecture. UniPocket uses frozen per-residue ESM-2 embeddings as input to a lightweight residual MLP with two task-specific heads, trained on ligand-contact labels from recently deposited PDB structures and cryptic-pocket annotations from CryptoBench. Training alternates between ligand-labeled and cryptic-labeled mini-batches, while an orthogonality regularizer encourages the two heads to learn complementary rather than redundant signals. UniPocket achieves a macro ROC-AUC of 0.82 for cryptic-pocket prediction and 0.76 for ligand-binding-site prediction - matching or exceeding all dedicated single-task baselines on both tasks simultaneously, without any 3D structural input at inference time.
Unsupervised dimensionality reduction methods aim to preserve intrinsic data geometry by maintaining local neighborhoods and approximate global relationships in low-dimensional embeddings, but they do not use label information and therefore may fail to reflect task-relevant class structure in biomedical and health applications. Supervised dimensionality reduction (SDR) incorporates labels to improve class organization, yet existing approaches often face a trade-off between discrimination and geometric faithfulness. Linear supervised methods are stable and interpretable but are limited in their ability to capture nonlinear structure, whereas many nonlinear methods impose supervision directly in the embedding space, which can over-separate classes and distort the underlying manifold. In biomedical applications, labels such as cell types in single-cell data or patient status in clinical cohorts provide meaningful biological signal, and supervised dimensionality reduction can use this information to produce more informative low-dimensional representations. Here we propose a new framework, DG-UMAP (Decision-Geometry UMAP), for faithful supervised dimensionality reduction via decision geometry. We first fit a classifier in the original feature space and use its boundary-local decision geometry to construct a low-rank metric deformation that emphasizes discriminative directions while limiting geometric distortion. Parametric UMAP is then applied to the transformed space, so supervision acts through the ambient geometry rather than by directly forcing class separation in the embedding. Across synthetic and multiple real-world biomedical datasets, our method yields embeddings with improved agreement with class structure and global organization while preserving local neighborhood quality.
DeepSSETracer is a method for segmenting protein secondary structure from medium-resolution (5-10Å) cryogenic electron microscopy (cryo-EM) density maps. We conducted experiments and ablation studies to examine the effects of normalization methods, max-pooling, activation functions, and loss calculation region on DeepSSETracer. By combining multiple technical improvements, the performance of the new version, DeepSSETracer 2.0, was significantly enhanced compared to DeepSSETracer 1.1. On a set of 77 test cases, the weighted average per-voxel F1 score increased from 62.1% to 70.3% for helix detection, and from 47.8% to 62.5% for β-sheet detection. While each of the five modifications in the network enhanced the detection of both helices and β-sheets, the improvement on β-sheets was even more pronounced. The ablation studies show that the most enhanced accuracy comes from the replacement of batch normalization with instance normalization, which accounts for increased F1 scores by 3% (helix) and 6.3% (β-sheet). These results show that relatively modest network tuning can significantly improve segmentation, suggesting that further incremental gains remain possible within the U-Net deep learning architecture.
The number of reference genomes is rapidly increasing, thanks to advances in long-read sequencing and assembly. While these collections can improve the sensitivity and specificity of classification methods, they require highly efficient compressed indexes. K-mer-based approaches like Kraken 2 are efficient but limit the analysis to a fixed k-mer length. This is hard for the user to set ahead of time, and suboptimal settings can harm sensitivity and specificity. Methods that use compressed full-text indexes like SPUMONI2 and Cliffy lift this constraint, but are less efficient than k-mer-based tools. Further, these methods either cannot report a full listing of genomes where a match occurs, or cannot scale to large reference databases. We propose new methods and algorithms that use compressed full-text indexes to enable multi-class and taxonomic classification. Unlike past compressed-indexing methods for classification, ours uses the move structure, which is extremely fast thanks to its locality of reference. Our method, called Movi Color, augments the main table of the Movi index. Specifically, Movi Color assigns a "color" to each run of the Burrows-Wheeler Transform according to the subset of genomes from which the run suffixes originated. When the reference is highly repetitive, as is typical when indexing pangenomes or reference databases, only certain colors occur, creating opportunities to compress the index. For species-level classification, Movi Color achieves over 1.9× higher positive predictive value (PPV) and about 3× higher sensitivity than Kraken 2 and Metabuli. At the genus level, it achieves 75% higher PPV than Metabuli, and over 50% higher sensitivity compared to Kraken 2. Movi Color's read processing time is 7-20× faster than Metabuli and is a comparable to Kraken 2. Although Movi Color uses more memory than both Kraken 2 and Metabuli, its speed-accuracy trade-off makes it well-suited for real-time and high-throughput scenarios.
Multimodal learning significantly improves cancer survival prediction using multiple data types for each patient, including clinical data, pathological images, and genomic data. However, existing works often assume that each data type contributes equally, neglecting potential conflicting or unreliable information between different modalities by simply concatenating features to form a fused representation. In addition, current frameworks are trained only to rank relative risk among patients and cannot quantify the actual likelihood of survival at specific time points, limiting their utility in real-world clinical settings. In this work, we propose the Modality-Aware Discrete-Time Survival (MADSurv) framework to address these gaps. First, it implements an uncertainty-aware attention mechanism in which modality-specific expert encoders learn both predictive features and their own confidence for intelligent data fusion. This approach leads to more robust and personalized predictions by dynamically focusing on the modalities that are most reliable for each individual patient. Second, in addition to a single overall risk score, MADSurv produces a sequence of survival probabilities for discrete yearly intervals. We evaluated our proposed method on five different cancer datasets. In addition to ranking patients' overall risk using the concordance index, we also assessed the accuracy of the model's survival probability estimates on each yearly milestone using the Brier score. Our experimental results demonstrate that MADSurv achieves superior and competitive performance compared to state-of-the-art methods.
Accurate quantifying dietary contents, such as calories, proteins, carbohydrates, and fats, from an image of a meal plate is vital for managing diabetes. Recently, Large Multimodal Models (LMMs) have excelled in complex vision-language tasks due to their use of very large, highly diverse data. This study benchmarked the use of seven LMMs that include full and lightweight models of GPT, Gemini, and Llama for nutrition estimation based on Google's Nutrition5k dataset and our own phone-collected DonateAndLearn dataset. We analyzed the performance of LMMs and the RGB-D fusion model, in which the RGB-D model was specifically trained using Nutrition5k data. On our DonateAndLearn dataset, the full-weight versions of LMMs significantly outperformed the RGB-D fusion model, suggesting superior generalization capacity of the LMMs. We propose a method to integrate Nutrition5k images with phone-collected meal images that often lack the physical sizes of objects in the images. We applied scaled phone images to the RGB-D fusion model to predict the total weight of food in each phone-collected image. Using the predicted weight, the mean absolute percentage error (MAPE) of carbohydrate prediction using the Gemini 2.5 Flash model decreased from 56.6% to 39.5%, based on 78 test cases in the DonateAndLearn dataset. Furthermore, when the ground-truth food weight was provided to Gemini2.5 Flash and GPT-4.1, the MAPE further improved dramatically to 20.2% and 26.8%, respectively, which underscores the critical value of integrating physical information into dietary assessment tools.
Hospital-acquired infections (HAIs) pose a significant challenge in healthcare settings, contributing substantially to patient morbidity, mortality, and increased healthcare costs. HAI incidence arises from complex interactions among healthcare workers, patients, and contaminated medical equipment. Agent-based modeling (ABM) is a well-known tool for simulating HAI dynamics, but existing ABM solutions have limited efficacy because of their reliance on synthetic/random human movement data. To address this gap, we deploy an Ultra-Wideband (UWB) Real-Time Location System (RTLS) in the post-surgery observation unit of a tertiary hospital in our city and collect high-resolution spatiotemporal location data of healthcare workers and medical devices in that unit. We develop an agent-based model of HAI transmission that incorporates epidemiological parameters specific to Clostridioides difficile (C. diff), capturing both direct and indirect transmission routes. The model is calibrated using assimilated RTLS data and is then applied to forecast exposure risk associated with asymptomatic carriers in the presence of biosecurity interventions (i.e., hand hygiene and surface disinfection). Our simulation results, generated in the AnyLogic simulation software, demonstrate that exposure levels vary due to movement behavior and infection control measures. These findings highlight the necessity of integrating real-time location data into ABMs to enhance predictive accuracy and optimize intervention strategies.
We present a privacy-preserving selection layer for collaborative population stratification under ϵ -local differential privacy (LDP). Rather than fixing a single pipeline (e.g., PCA+K-Means with preset K ), our framework lets parties choose among three DP pipelines: PCA→Noise, Noise→PCA, and Noise-Only, according to their resources, and has an honest-but-curious server aggregate only DP shares to automatically select the clustering algorithm (K-Means, GMM, or Hierarchical) and K that maximize internal metrics (Silhouette, Calinski-Harabasz, Davies-Bouldin). Because selection operates on DP data, it adds no further privacy loss. On openSNP (942 samples, 28,396 SNPs), the PCA-augmented pipelines yield higher utility and substantially lower communication and runtime than Noise-Only, and the recommended configuration consistently outperforms fixed baselines. Membership-inference attack power remains markedly lower for PCA-based pipelines across privacy budgets ϵ . In this paper, experiments are limited to two collaborating parties; extensions to multi-site collaboration are left for future work.
Early amyloid-β deposition is a hallmark of Alzheimer's disease (AD), though the exact nature of amyloid pathogenesis is not fully characterized. In this study, we designed a network diffusion model to simulate the spread of amyloid pathology through white matter brain networks of diagnostic subpopulations of healthy control (HC), mild cognitive impairment (MCI), and AD. Our network diffusion model was able to successfully model the spread of amyloid, recapturing regional distributions of amyloid observed in 18F-florbetapir positron emission tomography (r=0.44-0.46, P<0.01). When tuning the optimal parameters, we found that the optimal diffusion time (t) provided a notion of temporal progression, where the HC group had the lowest time (t = 107.22 ± 16.67), followed by MCI (t = 122.78 ± 19.63), and lastly AD (t =136.20 ± 24.47). The optimal starting seeds were the brainstem in all three diagnostic groups, followed by the lateral orbitofrontal lobes for HC and MCI and the lingual gyri in AD. Our findings corroborate evidence from amyloid staging studies where amyloid starts in the primary neocortex and associative cortex. The significance of the white matter structural network in the diffusion process provides evidence for the trans-synaptic spread hypothesis of amyloid in AD. In conclusion, our study provides novel insights into the pathogenesis of amyloid in AD and its subsequent propagation throughout the brain.
Early and accurate diagnosis of Alzheimer's disease (AD), a complex neurodegenerative disorder, requires analysis of heterogeneous biomarkers (e.g., neuroimaging, genetic risk factors, cognitive tests, and cerebrospinal fluid proteins) typically represented in a tabular format. With flexible few-shot reasoning, multimodal integration, and natural-language-based interpretability, large language models (LLMs) offer unprecedented opportunities for prediction with structured biomedical data. We propose a novel framework called TAP-GPT, Tabular Alzheimer's Prediction GPT, that adapts TableGPT2, a multimodal tabular-specialized LLM originally developed for business intelligence tasks, for AD diagnosis using structured biomarker data with small sample sizes. Our approach constructs few-shot tabular prompts using in-context learning examples from structured biomedical data and finetunes TableGPT2 using the parameter-efficient qLoRA adaption for a clinical binary classification task of AD or cognitively normal (CN). The TAP-GPT framework harnesses the powerful tabular understanding ability of TableGPT2 and the encoded prior knowledge of LLMs to outperform more advanced general-purpose LLMs and a tabular foundation model (TFM) developed for prediction tasks. To our knowledge, this is the first application of LLMs to the prediction task using tabular biomarker data, paving the way for future LLM-driven multi-agent frameworks in biomedical informatics.
Canonical correlation analysis (CCA) is a technique for finding correlations between different data modalities and learning low-dimensional representations. As fairness becomes crucial in machine learning, fair CCA has gained attention. However, previous approaches often overlook the impact on downstream classification tasks, limiting applicability. We propose a novel fair CCA method for fair representation learning, ensuring the projected features are independent of sensitive attributes, thus enhancing fairness without compromising accuracy. We validate our method on synthetic data and real-world data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), demonstrating its ability to maintain high correlation analysis performance while improving fairness in classification tasks. Our work enables fair machine learning in neuroimaging studies where unbiased analysis is essential. Code is available in https://github.com/ZhanliangAaronWang/FR-CCA-ADNI.
Imaging technologies have revolutionized the study of the tumor microenvironment (TME) by leveraging spatial analysis, which enables the exploration of tissue organization and cellular communication, as well as aiding cancer diagnosis and prognosis. However, while many advanced spatial analysis methods have been recently published, they are enmeshed with specific imaging technology. An opportunity exists to develop a technology-agnostic methodology that captures complex spatial patterns in the TME as phenotypes to use in downstream tasks. In this paper, we present a novel variation of spatial g-function and a comprehensive imaging-technologyagnostic framework that identifies rich spatial phenotypes that can be used in survival analysis and classification tasks. Applying our methodology to breast cancer, we uncover spatial phenotypes with significance to survival across racial groups and molecular subtypes of breast cancer. We find other phenotypes that are significant to the survival of specific patient categories (such as African American). We also demonstrate that our phenotypes reflect specific biological contexts. These results highlight the relevance of our proposed spatial analysis and phenotype discovery pipeline and demonstrate the benefits of the systematic exploration of spatial phenotypes for more personalized diagnosis and treatments.
Image registration is important in biological image analysis; however, it is often challenged by distortions and non-linear transformations. In this paper, we present a novel patch-wise image registration method to address the mentioned issues. Our method begins with global registration to correct linear transformations, followed by a detailed examination of geometrical distortions. After that, each image is adaptively divided into patches to isolate and correct non-linear distortions, followed by reconstruction and combining patches using Otsu thresholding. We evaluated our method against state-of-the-art techniques using mutual information (MI), phase congruency-based (PCB), and gradient-based metrics (GBM) across four real biology datasets. Our results demonstrate superior feature alignment and image coherence, especially in serial-stack registrations. While the proposed method has longer processing times compared to linear registration methods, its enhanced accuracy and reliability to handle non-uniform distortion makes it beneficial for precision-demanding applications. We have created a public GitHub repository containing the code used in our research, available at https://github.com/NabaviLab/CAPTURE.
Recently, there has been a revived interest in system neuroscience causation models due to their unique capability to unravel complex relationships in multi-scale brain networks. In this paper, our goal is to verify the feasibility and effectiveness of using a causality-based approach for fMRI fingerprinting. Specifically, we propose an innovative method that utilizes the causal dynamics activities of the brain to identify the unique cognitive patterns of individuals (e.g., subject fingerprint) and fMRI tasks (e.g., task fingerprint). The key novelty of our approach stems from the development of a two-timescale linear state-space model to extract 'spatio-temporal' (aka causal) signatures from an individual's fMRI time series data. To the best of our knowledge, we pioneer and subsequently quantify, in this paper, the concept of 'causal fingerprint.' Our method is well-separated from other fingerprint studies as we quantify fingerprints from a cause-and-effect perspective, which are then incorporated with a modal decomposition and projection method to perform subject identification and a GNN-based (Graph Neural Network) model to perform task identification. Finally, we show that the experimental results and comparisons with non-causality-based methods demonstrate the effectiveness of the proposed methods. We visualize the obtained causal signatures and discuss their biological relevance in light of the existing understanding of brain functionalities. Collectively, our work paves the way for further studies on causal fingerprints with potential applications in both healthy controls and neurodegenerative diseases.
The unprecedented scale of the COVID-19 pandemic created an alarming shortage of healthcare resources. To enable a more efficient resource allocation and targeted treatment, in this manuscript, we conducted a data-driven study of COVID-19 patients to predict patient outcomes and identify patient phenotypes. Specifically, we developed a multi-layered gated recurrent units-based model, referred to as mGRU-CP, to learn patient embeddings and estimate patient survival probabilities by leveraging their electronic health record (EHR) data in the COVID-19 Research Data Commons. We empirically compared mGRU-CP against four state-of-the-art baseline methods on three sets of patient features. The experimental results demonstrate that mGRU-CP could achieve competitive or superior performance over the baseline methods in all the settings. Our analysis also shows that the learned patient embeddings in mGRU-CP could enable meaningful patient phenotyping to better understand patient mortalities. Our study is significant in understanding patients in the past COVID-19 pandemic, and provides computational tools to predict patient outcomes and inform associated healthcare resource allocation for the future pandemics proactively.
Stroke is a significant cause of mortality and morbidity, necessitating early predictive strategies to minimize risks. Traditional methods for evaluating patients, such as Acute Physiology and Chronic Health Evaluation (APACHE II, IV) and Simplified Acute Physiology Score III (SAPS III), have limited accuracy and interpretability. This paper proposes a novel approach: an interpretable, attention-based transformer model for early stroke mortality prediction. This model seeks to address the limitations of previous predictive models, providing both interpretability (providing clear, understandable explanations of the model) and fidelity (giving a truthful explanation of the model's dynamics from input to output). Furthermore, the study explores and compares fidelity and interpretability scores using Shapley values and attention-based scores to improve model explainability. The research objectives include designing an interpretable attention-based transformer model, evaluating its performance compared to existing models, and providing feature importance derived from the model.
The irreversible and progressive atrophy by Alzheimer's Disease resulted in continuous decline in thinking and behavioral skills. To date, CNN classifiers were widely applied to assist the early diagnosis of AD and its associated abnormal structures. However, most existing black-box CNN classifiers relied heavily on the limited MRI scans, and used little domain knowledge from the previous clinical findings. In this study, we proposed a framework, named as PINet, to consider the previous domain knowledge as a Privileged Information (PI), and open the black-box in the prediction process. The input domain knowledge guides the neural network to learn representative features and introduced intepretability for further analysis. PINet used a Transformer-like fusion module Privileged Information Fusion (PIF) to iteratively calculate the correlation of the features between image features and PI features, and project the features into a latent space for classification. The Pyramid Feature Visualization (PFV) module served as a verification to highlight the significant features on the input images. PINet was suitable for neuro-imaging tasks and we demonstrated its application in Alzheimer's Disease using structural MRI scans from ADNI dataset. During the experiments, we employed the abnormal brain structures such as the Hippocampus as the PI, trained the model with the data from 1.5T scanners and tested from 3T scanners. The F1-score showed that PINet was more robust in transferring to a new dataset, with approximatedly 2% drop (from 0.9471 to 0.9231), while the baseline CNN methods had a 29% drop (from 0.8679 to 0.6154). The performance of PINet was relied on the selection of the domain knowledge as the PI. Our best model was trained under the guidance of 12 selected ROIs, major in the structures of Temporal Lobe and Occipital Lobe. In summary, PINet considered the domain knowledge as the PI to train the CNN model, and the selected PI introduced both interpretability and generalization ability to the black box CNN classifiers.
Structure-based drug discovery aims to identify small molecules that can attach to a specific target protein and change its functionality. Recently, deep learning has shown great promise in generating drug-like molecules with specific biochemical features and conditioned with structural features. However, they usually fail to incorporate an essential factor: the underlying physics which guides molecular formation and binding in real-world scenarios. In this work, we describe a physics-guided deep generative model for new ligand discovery, conditioned not only on the binding site but also on physics-based features that describe the binding mechanism between a receptor and a ligand. The proposed hybrid model has been tested on large protein-ligand complexes and small host-guest systems. Using the top-N methodology, on average more than 75% of the generated structures by our hybrid model were stronger binders than the original reference ligand. All of them had higher ΔGbind (affinity) values than the ones generated by the previous state-of-the-art method by an average margin of 1.88 kcal/mol. The visualization of the top-5 ligands generated by the proposed physics-guided model and the reference deep learning model demonstrate more feasible conformations and orientations by the former. The future directions include training and testing the hybrid model on larger datasets, adding more relevant physics-based features, and interpreting the deep learning outcomes from biophysical perspectives.
Arterial hypertension is a major risk factor for heart failure and antihypertensives such as angiotensin converting enzyme (ACE) inhibitors and β-blockers are considered as its first-line treatment. Drug response prediction models designed to determine the most effective antihypertensive drug for a patient are hindered by the interpatient response variability. Although typically pharmacogenetic data have been used to investigate the association of genetic variants with the antihypertensive response, genome-wide association studies are currently expensive and the translation of genotype guided antihypertensive therapy to clinical practice is challenging. With the generation of electronic health records (EHR) data summarized over the patient's disease prognosis and interventions, it is still an underused resource for antihypertensive effectiveness studies in heart failure management. In this study, we first use the clinical events in the EHR related to the patient's hard clinical endpoints and biomarkers associated with the heart failure condition to design selection strategies that determine the antihypertensive effectiveness, then develop annotated corpora using the strategies and eventually evaluate supervised deep learning classifiers on the annotated data. We annotated the EHR sequences of approximately 9500 patients with binary labels corresponding to the drug effectiveness across two different antihypertensive classes and our trained classifier was able to obtain the best F1 performance of 0.97.
One of the challenges in RNA-Seq studies is finding subsets of genes that share a common mechanism of action or are associated with a regulon/pathway. Existing approaches often extract modules that reflect quantitative similarities (such as genes with correlated log-fold-changes) but do not adequately capture biological significance. In this work, we propose the Dual ICA methodology, which provides an agnostic way to extract "interacting modules" composed of sets of genes and conditions that exhibit strong associations. Dual ICA involves performing Independent Component Analysis (ICA) twice, once on the genes and once on the conditions. Using the resulting signal matrices, we extract respective sets of genes and conditions. The interaction between these sets is quantified using the coefficients from a linear regression and significance is determined through the Wald test and Z-score filtering. These coefficients are equivalent to the outer product of independent components obtained from the two signal matrices. Not only do the gene sets extracted align with known regulons, but the significant interacting modules they instantiate also encompass conditions that influence the expression of these regulons through shared mechanisms of action. Compared to traditional unsupervised clustering methods, Dual ICA demonstrates superior performance and provides explicit gene-condition sets for exploring functional relationships.
